choice-metrics.com

Dear Ngene Team,

I am designing an experiment for four labeled alternatives (Petrol Car, Diesel Car, PHEV and BEV) with 8 attributes, which I will use for pilot survey to get parameter priors. Afterwards, I’m planning to collect data using Bayesian priors and estimate a panel MMNL model.

Here is my model specification followed by questions. The parameters in the model are all alternative-specific.

Code: Select all

design
;alts = PTRLC, DSLC, PHEV, BEV, NONE
;rows = 36
;block = 3
;eff = (mnl,d)

;model:

U(PTRLC) = b1[-0.0001]*price_P[12000,16000,20000] + b2[-0.0001]*fcost_P[1300,1700,2100]+ b3.dummy[-0.003|-0.005]*cazc_P[5,15,25]
         + b4.dummy[-0.0001]*cazhrs_P[0,1]+ b5[0.05]*nstations_P[50,100,150]+ b6[-0.01]*crtime_P[5,10,15]
         + b7[0.005]*range_P[400,700,1000] + b8[-0.0001]*cazc_P*cazhrs_P /

U(DSLC) = b9[0.001] + b10[-0.0001]*price_D[13000,17000,21000] + b11[-0.0001]*fcost_D[800,1200,1600]
        + b12.dummy[-0.003|-0.0001]*cazc_D[10,20,30]  + b13.dummy[-0.0002]*cazhrs_D[0,1] + b14[0.05]*nstations_D[50,100,150]
        + b15[-0.01]*crtime_D[5,10,15] + b16[0.005]*range_D[400,700,1000] + b17[-0.0001]*cazc_D*cazhrs_D /

U(PHEV) = b18[-0.05] + b19[-0.0002]*price_H[24400,32500,40600] + b20[-0.006]*fcost_H[400,700,1000]
          + b21[0.00001]*nstatiosn_H[50,100,150]  + b22[-0.001]*crtime_H[15,240,480] + b23[0.03]*range_H[50,500,1000]
          + b24[0.0003]*picg_H[2500,4500,6500] /

U(BEV) = b25[-0.05] + b26[-0.0001]*price_B[19500,26000,32500] + b27[-0.001]*fcost_B[300,500,700]+ b28[0.00001]*nstations_B[50,75,100]
        + b29[-0.001]*crtime_B[15,240,480] + b30[0.03]*range_B[100,150,200] + b31[0.0005]*picg_B[2500,4500,6500]  $

1. I used very small fixed parameters priors to get an initial efficient design. At this point, I can guess the signs of the parameters but the b and s-estimates do not seem to improve despite numerous tweaks and attempts. Is the design from this model good enough for a pilot survey in your eyes? Or am I better off with an orthogonal design?

2. Am I taking too much of a risk by making all the parameters alternative-specific? Most of the models I have seen in this forum
use generic-parameters even for labeled alternatives. I can make some of the parameters generic but not all given what I am interested in.

3.Some of the attributes are applicable to only two of the alternatives, for example CAZ charge is only applicable to petrol and diesel cars, not to PHEV and BEV. For this attribute, should I simply enter 0 (zero)or Not applicable (NA) for PHEV and BEV at the end of the design in ‘scenarios’?

Similarly, PiCG (grant) is only applicable to PHEV and BEV and is not applicable to Petrol and Diesel cars. In this case attribute PiCG will be entered as zero (0)in scenarios?

4. Can I estimate more two-way interaction effects after collecting the data, even though there are only two two-way interactions in this design?

5. I would like the size of the vehicle to pivot off the respondent's current car to account for hypothetical bias. The vehicle attributes in this design are for c-segment (medium cars) only. Could you give me an example from car choice studies of a pivot syntax for different car sizes?

6. To detect non-linear relationships, an attribute needs to have at least 3 levels but I often see dummy coded two-level attributes. What is the use of dummy coding two level attributes?

7. All my parameters are alternative specific, but just in case I change some of the parameters into generic parameters can I estimate market share and elasticity with labelled alternatives with generic parameters?

8. There are two other attributes I would like to use (CO2 emissions and NOx emissions) to capture respondent's views on environmental and air quality issues. But given that most DCE studies in the UK use lesser number of attributes, I am considering to you them as covariates. Would you advise me to increase my attributes from 8 to 10 or should I use them as covariates to minimise design and estimation complexity?

Sorry for asking too many questions. Your help is very much appreciated.

Kind regards,
Mike

Below my responses:

1. Using zero priors is a good way to create a design for a pilot study, there is no need to use an orthogonal design. Note that in a labelled experiment, dominant alternatives do not occur and as such Ngene is not using the sign of the priors to check for dominance. Therefore, you can simply use priors equal to zero. The priors you are using are actually quite far from zero, for example -0.0001 * 210000 = -2.1, which is very large. So you should use priors like 0.00000001 if you are using such large attribute level values, or in this case simply set to 0. Note that the S-estimate and B-estimate have no meaning in case of zero priors, so you can ignore them. You can focus on just the D-error and A-error.

2. Including alternative-specific parameters simply means that you need to estimate more coefficients, hence you need more information. This will only mean that you may need a larger sample size. But further there is no risk, optimising for a design with alternative-specific parameters will provide you with more flexibility and is better than optimising for generic parameters while estimating alternative-specific parameters.

3. You should only include attributes that are relevant in each alternative, so you should simply remove the other attributes from the utility specification, utility functions can be different for each labelled experiment.

4. Yes you can usually estimate many interaction effects, unless the interaction is perfectly correlated with another main or interaction effect (which is quite rare). Leaving them out will only not optimise the design for these interaction effects, but further there is usually no problem in model estimation.

5. I will answer this question in a separate post.

6. Dummy coding is useful for picking up nonlinearities in a continuous variable (like price), but dummy or effects coding is also necessary for categorical variable variables (which is different from nonlinear effects), such as brand names or colour. For example, when attribute levels are "blue" and "red", then there is no other way than to use dummy or effects coding. In case of a continuous variable with two levels there is no need to use dummy coding.

7. Yes you can always estimate market shares and elasticities if you have a labelled experiment, even if all your parameters are generic.

8. Attributes and covariates are two different things. Attributes are different across alternatives, while covariates are the same across alternatives (such as age and gender of the respondent). If you mean to include characterisstics of the respondent, then they need to be included as covariates, and will not be attributes on the choice experiment, you will simply add them later in the utility functions when estimating models.

Michiel

Responding the question 5:

You first need to determine whether you want to create a homogeneous design or a heterogeneous design. A homogeneous design is the same for all segments, while a heterogeneous design provides a different design for each segment.

1. A homogeneous design can only be created in Ngene if one of the alternatives is a status quo alternative with the reference levels of an average car in that segment. Since you are not presenting a status quo alternative, this option will not be useful in your case. You typically use syntax like:

U(SQ) = b1*x1.ref[10] + b2*x2.ref[2] /
U(alt1) = b1*x1.piv[-20%,0%,20%] + b2*x2.piv[-1,0,1,2] /
U(alt2) = b1*x1.piv[-20%,0%,20%] + b2*x2.piv[-1,0,1,2] /

2. Heterogeneous designs can be made by simply generating a separate design for each car segment based on the attribute levels of an average car in each segment. In other words, you create different versions of the survey. Namely, create a design for a specific segment (for example, for the medium class size, you can use your current syntax) and use this design for respondent with medium class cars. Similarly, generate designs for the other class sizes in a similar way. This way you generate a library of designs and you just look up the appropriate design for each respondent.

I hope this answers your question.

Hi Michiel,

Thank you very much for your quick answers to my questions. They are very useful. Just a quick question on your response to question 3:

I included only the attributes that are relevant to each alternative in the utility specification as you said. I use CAZ-charge as an attribute only for Petrol and Diesel car utility specifications. I understand that for this specific attribute, there will not be trade-offs across all four alternatives but two. However, I assume there will be respondents who, having made trade-offs across all the attributes - including CAZ charge- in a choice task, who will consider purchasing a plug-in hybrid or a BEV.

My question is, having collected the data, can I get information on how much a CAZ charge influences an individual's choice across all 4 alternatives, even though the charge is applicable only to Petrol and Diesel Cars?

On Qn 8, I framed the question somewhat incorrectly. What I meant to say was I should probably use a likert scale to capture respondents' views towards CO2 and NOx emissions prior to the choice tasks and use them as covariates? I was thinking to use emission levels from the cars as attributes and see how much they influence choice, but then again 10 attributes could be a bit too much? Thanks again.

Mike

Although CAZ charge is not an attribute in all alternatives, a change in level changes the choice probabilities of all alternatives, and hence captures the impact on all alternatives.

Attitudes towards emissions can be included in the utility function after data collection as you propose. I believe that the proper way to include attitudes with Likert scales in utility functions is through a hybrid choice model. This is however not an easy model to estimate, so I assume you will simply add them as a covariate.

Thank you, Michiel.

That answers all my questions for now Have a good day!

Mike